1/* 2 * Copyright © 2009 Intel Corporation 3 * 4 * Permission is hereby granted, free of charge, to any person obtaining a 5 * copy of this software and associated documentation files (the "Software"), 6 * to deal in the Software without restriction, including without limitation 7 * the rights to use, copy, modify, merge, publish, distribute, sublicense, 8 * and/or sell copies of the Software, and to permit persons to whom the 9 * Software is furnished to do so, subject to the following conditions: 10 * 11 * The above copyright notice and this permission notice (including the next 12 * paragraph) shall be included in all copies or substantial portions of the 13 * Software. 14 * 15 * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR 16 * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, 17 * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL 18 * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER 19 * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, 20 * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE 21 * SOFTWARE. 22 * 23 * Author: 24 * Zou Nan hai <nanhai.zou@intel.com> 25 * Yan Li <li.l.yan@intel.com> 26 * Liu Xi bin<xibin.liu@intel.com> 27 */ 28/* GRF allocation: 29 g1~g30: constant buffer 30 g1~g2:intra IQ matrix in UB format 31 g3~g4:non intra IQ matrix in UB format 32 g5~g20:IDCT table 33 g56~g79:DCT data after IQ before idct 34 g83~g106: IDCT data after idct 35 g82: thread payload backup 36 g125: ip before idct 37*/ 38IDCT_START: 39mov (1) g126.0<1>UD ip {align1}; 40jmpi DO_IDCT; 41add (16) g32<1>D g32<8,8,1>D ROW_ADD {compr}; 42add (16) g34<1>D g34<8,8,1>D ROW_ADD {compr}; 43add (16) g36<1>D g36<8,8,1>D ROW_ADD {compr}; 44add (16) g38<1>D g38<8,8,1>D ROW_ADD {compr}; 45 46shr (16) g32<1>D g32<8,8,1>D ROW_SHIFT {compr}; 47shr (16) g34<1>D g34<8,8,1>D ROW_SHIFT {compr}; 48shr (16) g36<1>D g36<8,8,1>D ROW_SHIFT {compr}; 49shr (16) g38<1>D g38<8,8,1>D ROW_SHIFT {compr}; 50 51mov (16) g110.0<1>W g32<16,8,2>W {align1}; 52mov (16) g111.0<1>W g34<16,8,2>W {align1}; 53mov (16) g112.0<1>W g36<16,8,2>W {align1}; 54mov (16) g113.0<1>W g38<16,8,2>W {align1}; 55 56mov (1) g80.0<1>UD a0.0<1,1,1>UD {align1}; //save a0 57mov (1) a0.0<1>UD 0x0DB00DA0UD {align1}; //begin at g110.0, the output of idct_row.g4i 58mov (1) g126.0<1>UD ip {align1}; 59jmpi DO_IDCT; 60 61add (16) g32<1>D g32<8,8,1>D COL_ADD {compr}; 62add (16) g34<1>D g34<8,8,1>D COL_ADD {compr}; 63add (16) g36<1>D g36<8,8,1>D COL_ADD {compr}; 64add (16) g38<1>D g38<8,8,1>D COL_ADD {compr}; 65 66shr (16) g32<1>D g32<8,8,1>D COL_SHIFT {compr}; 67shr (16) g34<1>D g34<8,8,1>D COL_SHIFT {compr}; 68shr (16) g36<1>D g36<8,8,1>D COL_SHIFT {compr}; 69shr (16) g38<1>D g38<8,8,1>D COL_SHIFT {compr}; 70 71mov (1) a0.0<1>UD g80.0<1,1,1>UD {align1}; //restore a0 72add (1) ip g125.0<1,1,1>UD 0x20UD {align1}; //jump back 73 74DO_IDCT: 75add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00200020UD {align1}; //increase the address 76dp4 (16) g40<1>D g[a0.0]<8,8,1>W g5<8,8,1>D {align1 compr}; 77dp4 (16) g42<1>D g[a0.0]<8,8,1>W g7<8,8,1>D {align1 compr}; 78dp4 (16) g44<1>D g[a0.0]<8,8,1>W g9<8,8,1>D {align1 compr}; 79dp4 (16) g46<1>D g[a0.0]<8,8,1>W g11<8,8,1>D {align1 compr}; 80dp4 (16) g48<1>D g[a0.0]<8,8,1>W g13<8,8,1>D {align1 compr}; 81dp4 (16) g50<1>D g[a0.0]<8,8,1>W g15<8,8,1>D {align1 compr}; 82dp4 (16) g52<1>D g[a0.0]<8,8,1>W g17<8,8,1>D {align1 compr}; 83dp4 (16) g54<1>D g[a0.0]<8,8,1>W g19<8,8,1>D {align1 compr}; 84add (2) g32.0<1>D g40.0<8,1,8>D g40.16<8,1,8>D {align1}; 85add (2) g33.0<1>D g42.0<8,1,8>D g42.16<8,1,8>D {align1}; 86add (2) g34.0<1>D g44.0<8,1,8>D g44.16<8,1,8>D {align1}; 87add (2) g35.0<1>D g46.0<8,1,8>D g46.16<8,1,8>D {align1}; 88add (2) g36.0<1>D g48.0<8,1,8>D g48.16<8,1,8>D {align1}; 89add (2) g37.0<1>D g50.0<8,1,8>D g50.16<8,1,8>D {align1}; 90add (2) g38.0<1>D g52.0<8,1,8>D g52.16<8,1,8>D {align1}; 91add (2) g39.0<1>D g54.0<8,1,8>D g54.16<8,1,8>D {align1}; 92 93add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00200020UD {align1}; 94dp4 (16) g40<1>D g[a0.0]<8,8,1>W g5<8,8,1>D {align1 compr}; 95dp4 (16) g42<1>D g[a0.0]<8,8,1>W g7<8,8,1>D {align1 compr}; 96dp4 (16) g44<1>D g[a0.0]<8,8,1>W g9<8,8,1>D {align1 compr}; 97dp4 (16) g46<1>D g[a0.0]<8,8,1>W g11<8,8,1>D {align1 compr}; 98dp4 (16) g48<1>D g[a0.0]<8,8,1>W g13<8,8,1>D {align1 compr}; 99dp4 (16) g50<1>D g[a0.0]<8,8,1>W g15<8,8,1>D {align1 compr}; 100dp4 (16) g52<1>D g[a0.0]<8,8,1>W g17<8,8,1>D {align1 compr}; 101dp4 (16) g54<1>D g[a0.0]<8,8,1>W g19<8,8,1>D {align1 compr}; 102add (2) g32.8<1>D g40.0<8,1,8>D g40.16<8,1,8>D {align1}; 103add (2) g33.8<1>D g42.0<8,1,8>D g42.16<8,1,8>D {align1}; 104add (2) g34.8<1>D g44.0<8,1,8>D g44.16<8,1,8>D {align1}; 105add (2) g35.8<1>D g46.0<8,1,8>D g46.16<8,1,8>D {align1}; 106add (2) g36.8<1>D g48.0<8,1,8>D g48.16<8,1,8>D {align1}; 107add (2) g37.8<1>D g50.0<8,1,8>D g50.16<8,1,8>D {align1}; 108add (2) g38.8<1>D g52.0<8,1,8>D g52.16<8,1,8>D {align1}; 109add (2) g39.8<1>D g54.0<8,1,8>D g54.16<8,1,8>D {align1}; 110 111add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00200020UD {align1}; 112dp4 (16) g40<1>D g[a0.0]<8,8,1>W g5<8,8,1>D {align1 compr}; 113dp4 (16) g42<1>D g[a0.0]<8,8,1>W g7<8,8,1>D {align1 compr}; 114dp4 (16) g44<1>D g[a0.0]<8,8,1>W g9<8,8,1>D {align1 compr}; 115dp4 (16) g46<1>D g[a0.0]<8,8,1>W g11<8,8,1>D {align1 compr}; 116dp4 (16) g48<1>D g[a0.0]<8,8,1>W g13<8,8,1>D {align1 compr}; 117dp4 (16) g50<1>D g[a0.0]<8,8,1>W g15<8,8,1>D {align1 compr}; 118dp4 (16) g52<1>D g[a0.0]<8,8,1>W g17<8,8,1>D {align1 compr}; 119dp4 (16) g54<1>D g[a0.0]<8,8,1>W g19<8,8,1>D {align1 compr}; 120add (2) g32.16<1>D g40.0<8,1,8>D g40.16<8,1,8>D {align1}; 121add (2) g33.16<1>D g42.0<8,1,8>D g42.16<8,1,8>D {align1}; 122add (2) g34.16<1>D g44.0<8,1,8>D g44.16<8,1,8>D {align1}; 123add (2) g35.16<1>D g46.0<8,1,8>D g46.16<8,1,8>D {align1}; 124add (2) g36.16<1>D g48.0<8,1,8>D g48.16<8,1,8>D {align1}; 125add (2) g37.16<1>D g50.0<8,1,8>D g50.16<8,1,8>D {align1}; 126add (2) g38.16<1>D g52.0<8,1,8>D g52.16<8,1,8>D {align1}; 127add (2) g39.16<1>D g54.0<8,1,8>D g54.16<8,1,8>D {align1}; 128 129add (1) a0.0<1>UD a0.0<1,1,1>UD 0x00200020UD {align1}; 130dp4 (16) g40<1>D g[a0.0]<8,8,1>W g5<8,8,1>D {align1 compr}; 131dp4 (16) g42<1>D g[a0.0]<8,8,1>W g7<8,8,1>D {align1 compr}; 132dp4 (16) g44<1>D g[a0.0]<8,8,1>W g9<8,8,1>D {align1 compr}; 133dp4 (16) g46<1>D g[a0.0]<8,8,1>W g11<8,8,1>D {align1 compr}; 134dp4 (16) g48<1>D g[a0.0]<8,8,1>W g13<8,8,1>D {align1 compr}; 135dp4 (16) g50<1>D g[a0.0]<8,8,1>W g15<8,8,1>D {align1 compr}; 136dp4 (16) g52<1>D g[a0.0]<8,8,1>W g17<8,8,1>D {align1 compr}; 137dp4 (16) g54<1>D g[a0.0]<8,8,1>W g19<8,8,1>D {align1 compr}; 138add (2) g32.24<1>D g40.0<8,1,8>D g40.16<8,1,8>D {align1}; 139add (2) g33.24<1>D g42.0<8,1,8>D g42.16<8,1,8>D {align1}; 140add (2) g34.24<1>D g44.0<8,1,8>D g44.16<8,1,8>D {align1}; 141add (2) g35.24<1>D g46.0<8,1,8>D g46.16<8,1,8>D {align1}; 142add (2) g36.24<1>D g48.0<8,1,8>D g48.16<8,1,8>D {align1}; 143add (2) g37.24<1>D g50.0<8,1,8>D g50.16<8,1,8>D {align1}; 144add (2) g38.24<1>D g52.0<8,1,8>D g52.16<8,1,8>D {align1}; 145add (2) g39.24<1>D g54.0<8,1,8>D g54.16<8,1,8>D {align1}; 146 147add (1) ip g126.0<1,1,1>UD 0x20UD {align1}; //jump back 148